Should You Trust ChatGPT With Your Health?
Karan Singhal of OpenAI argues that a general model with specialized health training, expert-calibrated evaluation, and partners handling the last mile can make trustworthy health intelligence abundant, while humans stay central to care.
Watch on YouTube ↗Charlie's summary
- ▶ 6:07
Specialized health training inside a general frontier model beats a separate specialized healthcare model
Karan, who used to build specialized models, says OpenAI adds specialized healthcare data to every phase of training of its ChatGPT models. A smaller healthcare-only model “would just kind of fall behind the frontier” quickly because general models improve so fast. OpenAI instead hitches a ride on scale while integrating evals and physician-sourced data.
- ▶ 2:32
AI's most useful role is between visits, as a safety net for patients and clinicians
Karan says people use ChatGPT to prepare for the roughly 15-minute visit, follow up afterward, and understand results. He is excited about a proactive agent, such as flagging a rising A1C for medication review. In OpenAI's study with Penda Health, a clinician copilot caught issues like medication interactions. He does not see it as a replacement for a doctor.
- ▶ 17:19
Aggregate many physicians, because five doctors give eight answers, and prioritize safety over reassurance
Karan says relying on a few in-house clinicians is not the best approach: “if you ask five physicians... you'll get like eight different answers.” OpenAI built a global network of hundreds of experts across dozens of specialties and languages. When a user wants comfort but the situation is unsafe, evaluations check that the model says to seek care. The approach has produced dramatic improvements.
- ▶ 20:22
AI can shrink the 5-to-17-year lag before clinical guidelines reach everyday practice
Karan describes the pipeline from papers to guidelines to practice, and says studies found guidelines take 5 to 17 years to penetrate practice. Bringing the latest literature into the point of care, as in ChatGPT for clinicians, aims to compress that to weeks, days or hours. For real uncertainty, models should recognize and hedge, which HealthBench evaluates.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.